Deep learning models for cell cycle phase prediction from single-cell RNA sequencing data.

Area of Science:

• Genomics
• Computational Biology
• Biotechnology

Background:

• Accurate cell cycle phase prediction is crucial for single-cell RNA sequencing (scRNA-seq) analysis, impacting disease research like cancer.
• Existing methods require evaluation for robustness and scalability across diverse biological samples.

Purpose of the Study:

• To evaluate traditional and deep learning models for cell cycle phase prediction using scRNA-seq data.
• To assess model performance across various datasets with experimentally verified labels.

Main Methods:

• Trained and evaluated machine learning (AdaBoost, Random Forest, LightGBM) and deep learning (DNNs, CNNs, ensembles) models.
• Utilized consensus phase labels from four tools (CellCycleScore, ccAFv2, Revelio, Tricycle) on diverse scRNA-seq datasets.
• Validated performance on independent datasets with experimentally verified cell cycle labels.

Main Results:

• Deep learning models, particularly ensembles, demonstrated strong cross-dataset prediction accuracy.
• REH-trained models achieved up to 74.35% accuracy on independent datasets.
• SHapley Additive exPlanations (SHAP) confirmed cell-cycle-related gene importance; lower marker recovery reduced performance.

Conclusions:

• Deep learning offers a robust and scalable solution for cell cycle phase prediction from scRNA-seq data.
• Model performance is influenced by the quality and completeness of cell cycle marker information in the training data.